Ultralytics Detect3D evaluation artifacts
Research evaluation artifacts for the native Ultralytics Detect3D implementation in ultralytics/ultralytics#25629.
Primary results
Primary checkpoints use the standard Chen KITTI train/validation split (3,712/3,769 images). KITTI R40 AP3D uses class-standard IoU thresholds. Latency is batch-1 FP32 inference at 416x1280 on an RTX 4090 after warmup, excluding preprocessing and postprocessing.
| Model | Params | GFLOPs | Car mod. AP3D | Pedestrian mod. AP3D | Cyclist mod. AP3D | PT latency | ONNX CUDA latency |
|---|---|---|---|---|---|---|---|
| YOLO26n-3d, seed 0 | 3.12M | 8.1 | 9.67 | 4.65 | 5.37 | 5.925 ms | 2.765 ms |
| YOLO26s-3d, seed 1 | 10.73M | 25.1 | 13.18 | 6.06 | 5.10 | 6.694 ms | 3.017 ms |
Full Chen-split KITTI R40 results
Values are Easy / Moderate / Hard percentages.
| Model | Class | AP3D | APBEV | AOS |
|---|---|---|---|---|
| YOLO26n-3d, seed 0 | Car | 12.87 / 9.67 / 8.13 | 21.18 / 15.76 / 13.58 | 94.23 / 86.18 / 79.73 |
| YOLO26n-3d, seed 0 | Pedestrian | 5.84 / 4.65 / 3.69 | 6.79 / 5.52 / 4.51 | 50.26 / 42.76 / 37.91 |
| YOLO26n-3d, seed 0 | Cyclist | 10.40 / 5.37 / 4.92 | 10.88 / 5.61 / 5.37 | 41.89 / 27.25 / 26.11 |
| YOLO26s-3d, seed 1 | Car | 17.95 / 13.18 / 11.21 | 26.30 / 19.66 / 17.03 | 96.40 / 90.61 / 82.69 |
| YOLO26s-3d, seed 1 | Pedestrian | 8.28 / 6.06 / 4.91 | 9.96 / 7.25 / 5.99 | 56.33 / 47.84 / 41.46 |
| YOLO26s-3d, seed 1 | Cyclist | 10.31 / 5.10 / 4.96 | 10.96 / 5.72 / 5.39 | 47.60 / 31.24 / 30.03 |
Supplementary 80/20 results
These checkpoints use a deterministic stratified 80/20 split (5,985/1,496 images). They must not be compared directly with the Chen split or external KITTI results because correlated frames can cross the image-level split.
Full 80/20 KITTI R40 results
Values are Easy / Moderate / Hard percentages.
| Model | Class | AP3D | APBEV | AOS |
|---|---|---|---|---|
| YOLO26n-3d, seed 2 | Car | 65.57 / 51.89 / 46.44 | 74.00 / 58.89 / 53.25 | 98.60 / 98.03 / 95.36 |
| YOLO26n-3d, seed 2 | Pedestrian | 24.60 / 19.84 / 17.69 | 27.09 / 22.01 / 19.80 | 89.25 / 84.34 / 81.18 |
| YOLO26n-3d, seed 2 | Cyclist | 35.11 / 28.54 / 27.68 | 38.47 / 30.94 / 30.20 | 89.20 / 87.40 / 85.09 |
| YOLO26s-3d, seed 0 | Car | 75.05 / 60.84 / 54.68 | 81.15 / 67.45 / 61.28 | 99.18 / 99.06 / 96.52 |
| YOLO26s-3d, seed 0 | Pedestrian | 30.53 / 24.46 / 21.13 | 32.19 / 25.77 / 22.34 | 94.11 / 88.52 / 85.19 |
| YOLO26s-3d, seed 0 | Cyclist | 42.91 / 36.41 / 34.66 | 45.39 / 38.43 / 36.54 | 91.87 / 91.44 / 91.05 |
Visualizations
Projected 3D predictions from the primary Chen-split YOLO26s-3d checkpoint:
KITTI R40 summary and training curves
Assets and provenance
Primary checkpoints:
chen-yolo26n-3d-seed0.ptand.onnxchen-yolo26s-3d-seed1.ptand.onnx
Supplementary checkpoints:
kitti80-20-yolo26n-3d-seed2.ptand.onnxkitti80-20-yolo26s-3d-seed0.ptand.onnx
The curves/logs archive contains recorded training arguments, epoch metrics, fresh validation curves, KITTI R40 plots, and the consolidated validation log. Raw RTX 4090 speed measurements, complete generic metrics, training commands, data split details, and checksums are attached separately.
The archived args.yaml files preserve the original training provenance and include experimental keys that are not accepted by the final release code. Use the commands in detect3d-best-seed-summary.md instead of replaying those YAML files directly. The ONNX models use opset 17 with static 1x3x416x1280 input and 1x300x14 output; re-export from PT for other shapes.
These checkpoints were trained with an earlier contributor revision and revalidated/exported with the final PR code. They are contributor research artifacts, not official Ultralytics weights. KITTI data is not redistributed; use is subject to the KITTI dataset terms and applicable Ultralytics licensing. Verify downloads with SHA256SUMS.